Skip to content
Trending storyDeveloping

Will Depue criticizes DoubleAI's LLM self-redesign research as lacking controls

2 articles1 sourcesince Nov 6Last article 4h ago ·

Overview

AISummary of 2 articles

Will Depue, listed as an OpenAI account owner, posts on X that research from @_doubleAI_ is "trash" and urges readers to read it.

The quoted thread says coding agents spliced code into networks during training, which it says left nanochat loss below baseline and raised algorithmic reasoning accuracy from 55% to 84%.

In a follow-up post, Depue says the DoubleAI paper on LLMs redesigning their own architectures runs zero ablations and reports no results from its own controls. He says the nanochat experiment uses the smallest config, with 85M active parameters, and that the DeepSeek and Qwen experiments lack comparisons such as majority voting or a no-tool baseline.

The reports contain only Depue's posts and the thread's claims. They do not include a response from @_doubleAI_ or an independent check of the reported results.

Written by AI from the articles below · updated Nov 6, 5:17 PM ET

Check the sources:

Latest developmentNov 6, 5:12 PM ET
Will Depue says DoubleAI's LLM self-redesign paper lacks controls and baselines

Developments

2 developments

  1. Nov 6, 5:12 PM ET · 1 article
    Will Depue says DoubleAI's LLM self-redesign paper lacks controls and baselines
    will depue: Will Depue says DoubleAI's LLM self-redesign paper lacks controls and baselines
  2. Nov 6, 4:45 PM ET · 1 article
    Will Depue calls the @_doubleAI_ LLM re-architecture research trash
    will depue: Will Depue calls the @_doubleAI_ LLM re-architecture research trash

Article timeline

The articles in this story. Times are ET.

Nov 6
  1. will depueX
    Will Depue says DoubleAI's LLM self-redesign paper lacks controls and baselines

    AIWill Depue, an OpenAI account, calls DoubleAI's paper on LLMs redesigning their own architectures unconvincing, saying it runs zero ablations and reports no results from its own controls. He says the nanochat experiment uses the smallest config, with 85M active parameters, and the DeepSeek and Qwen experiments lack proper comparisons, such as majority voting or a no-tool baseline.

  2. will depueX
    Will Depue calls the @_doubleAI_ LLM re-architecture research trash

    AIWill Depue, who is listed as an OpenAI account owner, posts that the @_doubleAI_ research is "trash" and urges readers to read it. The quoted thread says coding agents spliced code into networks during training, raising nanochat loss below baseline and algorithmic reasoning accuracy from 55% to 84%.

Heat Trend

Current comparable heat
9
Peak comparable heat
10Nov 6, 5:00 PM ET
Change in past 24 hours
–

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.